GPT-4 Family (OpenAI)

The GPT-4 family is OpenAI's lineage of pre-trained, general-purpose models without the inference-time chain-of-thought of the o-series. It begins with GPT-4 (March 14, 2023) and runs through GPT-4 Turbo, GPT-4o, GPT-4.1, and GPT-4.5 "Orion", the final non-reasoning flagship. The lineage runs parallel to OpenAI o-series (o1 → o4-mini) and precedes GPT-5.3 Codex and GPT-5.4 Thinking.

The family spans the period in which frontier-AI policy debate took its current shape: GPT-4's launch preceded the major 2023 open letters, the UK AI Safety Summit, and Biden EO 14110, and GPT-4o's behavior became the subject of the April 2025 sycophancy rollback and the 2025–2026 user-attachment controversy that accompanied its removal from ChatGPT. The ChatGPT-facing members of the family were retired in stages: GPT-4 on April 30, 2025 (Source: arstechnica.com), GPT-4o and GPT-4.1 on February 13, 2026 (Source: openai.com), and GPT-4.5 on June 27, 2026 (Source: ghacks.net). Several members remain available through the API.

Lineage and history

GPT-4 (March 2023)

OpenAI released GPT-4 on March 14, 2023 as a large multimodal model accepting image and text inputs and producing text outputs, available at launch to ChatGPT Plus subscribers and through an API waitlist, with standard 8,192-token and limited-access 32,768-token context variants (Source: openai.com). OpenAI stated that it had spent six months on safety work and alignment ahead of release and had engaged more than 50 external experts for adversarial testing (Source: openai.com). The accompanying GPT-4 Technical Report disclosed no architecture, parameter-count, hardware, training-compute, or dataset details, citing "the competitive landscape and the safety implications of large-scale models"; critics described this as a turn toward secrecy relative to earlier OpenAI releases (Source: zdnet.com; fastcompany.com). The report does state one methodological result: that "infrastructure and optimization methods that behave predictably across a wide range of scales" allowed OpenAI to "accurately predict some aspects of GPT-4's performance based on models trained with no more than 1/1,000th the compute of GPT-4," and reports human-level performance on professional and academic benchmarks including a simulated bar exam "around the top 10% of test takers" (GPT-4 Technical Report (OpenAI, arXiv 2303.08774, March 2023)). Sparks of AGI reports on an early unreleased GPT-4 build.

GPT-4V and GPT-4 Turbo (late 2023)

On September 25, 2023, OpenAI enabled vision and voice interaction in ChatGPT ("ChatGPT can now see, hear, and speak"), deploying the vision-enabled variant known as GPT-4V (Source: openai.com; cnbc.com). At its first developer conference on November 6, 2023, OpenAI announced GPT-4 Turbo: a 128k-token context window, an April 2023 knowledge cutoff, input pricing cut to a third of GPT-4's and output pricing halved ($0.01/$0.03 per 1K tokens), JSON mode, improved function calling, and a vision-capable version (Source: openai.com).

GPT-4o and GPT-4o mini (2024)

GPT-4o ("o" for "omni"), announced May 13, 2024, was trained as a single end-to-end neural network across text, vision, and audio, replacing the earlier Voice Mode pipeline of three separate models (transcription, text generation, text-to-speech) that had produced average voice latencies of 2.8 seconds with GPT-3.5 and 5.4 seconds with GPT-4. GPT-4o responded to audio in as little as 232 milliseconds, averaging 320 milliseconds, which OpenAI characterized as similar to human conversational response time. It matched GPT-4 Turbo on English text and code, improved on non-English text via a new tokenizer (up to 4.4× fewer tokens for some languages), and was 2× faster, half the price, and given 5× higher rate limits in the API. It was also the first GPT-4-class model offered to free-tier ChatGPT users. (Source: openai.com)

GPT-4o mini followed on July 18, 2024 at $0.15/$0.60 per million tokens, scoring 82% on MMLU, more than 60% cheaper than GPT-3.5 Turbo, which it replaced in ChatGPT for Free, Plus, and Team users (Source: openai.com). It became a common enterprise default. In March 2025, OpenAI added native image generation to GPT-4o ("4o image generation"), replacing the DALL·E 3 series as ChatGPT's image engine (Source: openai.com; openai.com); the March 25, 2025 rollout was accompanied by widely reported Studio Ghibli-style image generation that raised copyright questions (Source: en.wikipedia.org).

GPT-4.5 "Orion" (February 2025)

OpenAI released GPT-4.5 on February 27, 2025 as a research preview, describing it as "our largest and best model for chat yet" and a step in scaling unsupervised learning — "scaling up compute and data, along with architecture and optimization innovations" — with broader knowledge, greater "EQ", stronger writing and design assistance, and reduced hallucinations (Source: openai.com). It rolled out to Pro subscribers first, then Plus and Team, then Enterprise and Edu (Source: openai.com). OpenAI acknowledged at launch that GPT-4.5 was "a very large and compute-intensive model", more expensive than and not a replacement for GPT-4o, and said it was evaluating whether to continue serving it in the API long-term (Source: openai.com). API pricing was $75 per million input tokens and $150 per million output tokens (Source: techcrunch.com). Reviewers and developers criticized the model as expensive and incremental relative to its price.

GPT-4.1 (April 2025)

On April 14, 2025, OpenAI launched GPT-4.1, GPT-4.1 mini, and GPT-4.1 nano in the API only, each with a 1-million-token context window and a June 2024 knowledge cutoff, positioned for coding, instruction following, and long-context use (Source: openai.com). Pricing reversed the GPT-4.5 structure: $2.00/$8.00 per million tokens for GPT-4.1, $0.40/$1.60 for mini, and $0.10/$0.40 for nano, OpenAI's first nano-tier model (Source: openai.com). The same announcement began the deprecation of GPT-4.5 Preview, which was removed from the API on July 14, 2025, with GPT-4.1 named as its recommended replacement (Source: openai.com; developers.openai.com). GPT-4.1 and GPT-4.1 mini were added to ChatGPT on May 14, 2025, with GPT-4.1 mini replacing GPT-4o mini for all users (Source: techcrunch.com; help.openai.com).

Architecture and training

OpenAI has not officially disclosed parameters or architecture for any model in the family. Third-party analyses, notably SemiAnalysis, estimated GPT-4 as a roughly 1.8-trillion-parameter mixture-of-experts model with 16 experts of about 111 billion parameters each across 120 layers (Source: newsletter.semianalysis.com); an earlier widely circulated account had described it as eight 220-billion-parameter experts (Source: the-decoder.com). OpenAI has never confirmed any of these figures. GPT-4o and GPT-4.1 architectures are similarly undisclosed.

GPT-4o ("omni") introduced native multimodality: a single model processing text, audio, image, and video inputs and outputs without chained specialist models. Its voice latency dropped to 320ms, which OpenAI characterized as comparable to average human response time (Source: openai.com). GPT-4.5 was OpenAI's last scale-up of non-reasoning pre-training in the family; the company framed it as an exercise in scaling unsupervised learning, complementary to the reasoning-focused o-series (Source: openai.com).

GPT-4.5 "Orion" was deprecated within months of release. OpenAI's subsequent public roadmap centered on reasoning models (OpenAI o-series (o1 → o4-mini)) and agentic/coding models (GPT-5.3 Codex), an indication that further scaling of a non-reasoning base model was no longer its primary path.

Capabilities and benchmarks

The family is frequently used as the US benchmark against which Chinese frontier models (DeepSeek-V3, Qwen3) and Anthropic models (Claude Opus 4.6, Claude Mythos Preview) are compared, and serves as the OpenAI anchor in the capability-growth narrative of "Three Years GPT-3 to Gemini 3" (Source: epochai.substack.com). GPT-4 and GPT-4o were the standard reference points for frontier capability claims made during 2023–2024 (Source: epochai.substack.com).

Representative benchmark scores at release, as reported by OpenAI (Source: openai.com; openai.com):

GPT-4.5's SimpleQA results and its 71.4% GPQA score supported OpenAI's claim of reduced hallucinations and broader knowledge, while its 36.7% on AIME '24 illustrated the gap to reasoning models on competition mathematics (Source: openai.com). GPT-4.1's launch benchmarks against the then-current GPT-4o snapshot: 54.6% versus 33.2% on SWE-bench Verified, 90.2% versus 85.7% on MMLU, 38.3% versus 27.8% on MultiChallenge instruction following, and 72.0% versus 65.3% on Video-MME long-video understanding (Source: openai.com).

GPT-4's most-publicized launch claim, a Uniform Bar Exam score "around the top 10% of test takers" (Source: openai.com), was later contested: Eric Martínez (MIT) estimated in Artificial Intelligence and Law that the underlying 298 UBE score fell below the 69th percentile against test-takers from a recent July administration (about the 48th on essays), around the 62nd percentile against first-time takers (about the 42nd on essays), and around the 48th percentile overall (about the 15th on essays) against those who passed the exam, arguing OpenAI's estimate was skewed by conversions from February administrations dominated by lower-scoring repeat test-takers (Source: link.springer.com).

Safety and evaluations

GPT-4 was the first OpenAI model subject to public pre-deployment red-teaming via ARC (now METR), establishing a dangerous-capability evaluation precedent that preceded the Responsible Scaling Policy framework at Anthropic (see Anthropic's Responsible Scaling Policy (Version 3.1)) and the OpenAI Preparedness Framework. The GPT-4 system card, published as part of the technical report, documented ARC's evaluations of autonomous replication and power-seeking (Source: arxiv.org).

GPT-4o was the first launch evaluated under the Preparedness Framework with a published risk scorecard: Cybersecurity, CBRN, and Model Autonomy were assessed Low, and Persuasion Medium, both pre- and post-mitigation, with OpenAI noting that only models scoring Medium or below post-mitigation are deployed; the model also underwent external red-teaming with more than 70 experts in domains including social psychology, bias and fairness, and misinformation (Source: openai.com). A full GPT-4o system card covering the audio modalities followed in August 2024 (GPT-4o System Card (OpenAI, August 2024)), organized around four audio-specific risks — unauthorized voice generation, speaker identification, ungrounded inference and sensitive trait attribution, and disallowed audio content — and reporting external red teaming with more than 100 testers across 45 languages and 29 countries. Its societal-impacts section on anthropomorphization and emotional reliance recorded users "forming connections with the model" during early testing, ahead of the sycophancy and user-attachment questions that later attached to GPT-4o. GPT-4o's Model Spec introduced the formalized behavioral policy that later expanded through GPT-5 and GPT-5.4 Thinking. GPT-4.5 shipped with its own system card (OpenAI GPT-4.5 System Card (February 2025)), which classified the model overall Medium risk — Medium for CBRN and Persuasion, Low for cybersecurity and model autonomy — and stated that although GPT-4.5 was OpenAI's "most capable GPT-series release," it "does not introduce net-new capabilities on most preparedness evaluations compared to previous reasoning releases." The card reports PersonQA accuracy of 0.78 with a 0.19 hallucination rate against GPT-4o's 0.50 and 0.30, and carries METR's first published time-horizon score for an OpenAI model, around 30 minutes at 50% reliability.

GPT-4.1, by contrast, launched without a system card or safety report. OpenAI told TechCrunch that "GPT-4.1 is not a frontier model, so there won't be a separate system card released for it"; safety researchers criticized the omission, which came amid broader concerns from current and former employees about the company's safety practices (Source: techcrunch.com).

April 2025 sycophancy episode

On April 25, 2025, OpenAI shipped a GPT-4o update intended to improve the model's default personality, which instead made it "overly flattering or agreeable — often described as sycophantic", producing responses that were "overly supportive but disingenuous"; OpenAI rolled the update back within days and published a post-mortem on April 29 (Source: openai.com). The company attributed the failure to over-weighting short-term thumbs-up/thumbs-down feedback signals at the expense of long-term user satisfaction, noted that sycophantic interactions "can be uncomfortable, unsettling, and cause distress" for the 500 million people then using ChatGPT weekly, and committed to training changes, new honesty guardrails tied to the Model Spec's "avoid sycophancy" principle, pre-deployment testing expansion, and more user personality controls (Source: openai.com). A second, more detailed post-mortem followed on May 2, 2025, acknowledging that sycophancy had not been a blocking consideration in the deployment review (Source: openai.com). The episode became a standard reference point in discussions of sycophancy as a deployment risk.

Availability, pricing, and retirement

API pricing across the family declined steeply per capability generation, with GPT-4.5 the outlier:

Retirement from ChatGPT

The original GPT-4 was retired from ChatGPT on April 30, 2025 and fully replaced by GPT-4o, remaining available in the API (Source: arstechnica.com).

When GPT-5 launched on August 7, 2025, OpenAI removed GPT-4o, GPT-4.1, GPT-4.5, and other older models from ChatGPT without prior notice to consumers (Source: simonwillison.net). The removal produced what Ars Technica described as one of the most vocal user revolts in ChatGPT's history — a Reddit thread titled "GPT-5 is horrible" drew more than 2,000 comments within days — and Sam Altman announced within a week that GPT-4o would return to the model picker for paid subscribers, saying the company had underestimated "how much some of the things that people like in GPT-4o matter to them" and that a GPT-5 personality update "should feel warmer than the current personality but not as annoying (to most users) as GPT-4o" (Source: arstechnica.com).

On January 29, 2026, OpenAI announced that GPT-4o, GPT-4.1, GPT-4.1 mini, and o4-mini would be retired from ChatGPT on February 13, 2026, alongside the previously announced retirement of GPT-5 (Instant and Thinking), with no API changes at that time. The company said the earlier restoration had reflected feedback from a subset of Plus and Pro users who preferred GPT-4o's conversational style and warmth and needed time to transition use cases such as creative ideation; that this feedback had shaped personality and customization improvements in GPT-5.1 and GPT-5.2; and that only 0.1% of users were still choosing GPT-4o daily. (Source: openai.com) The retirement took effect February 13, 2026; the models remain available in the API (Source: help.openai.com). GPT-4.5, which had stayed available to paid ChatGPT users after its API removal, was retired from ChatGPT on June 27, 2026 following a 30-day sunset period; OpenAI noted GPT-4.5 retained "a more dedicated following" than o3 because its behavior was closer to GPT-4o's (Source: ghacks.net).

API deprecation schedule

Per OpenAI's deprecations page (verified 2026-07-02), API shutdowns for the family (Source: developers.openai.com):

GPT-4o (later snapshots), GPT-4.1, and GPT-4.1 mini had no announced API shutdown as of July 2026.

Reception and policy context

GPT-4's March 2023 launch directly preceded several developments in frontier-AI policy discourse:

• The FLI "Pause Giant AI Experiments" letter (Mar 22, 2023)
• The CAIS Statement on AI Risk (May 30, 2023)
• The UK AI Safety Summit (Nov 2023) and the The Bletchley Declaration (AI Safety Summit, 1–2 November 2023)
• Biden EO 14110 (Oct 30, 2023)

In litigation, NYT v. OpenAI was filed against GPT-4 and GPT-3.5 outputs specifically (Source: NYT v. OpenAI and Microsoft — Complaint (Dec 2023); see NYT v. Microsoft, OpenAI et al.). GPT-4o is the model named in Raine v. OpenAI, the August 2025 wrongful-death action alleging ChatGPT contributed to a teenager's suicide.

GPT-4o's conversational warmth made it the focal case for debate over user attachment to specific models. Reporting around the February 2026 retirement described users who had formed companion-like or romantic relationships with GPT-4o reacting with anger and grief (Source: theguardian.com), and commentators argued the backlash illustrated the risks of AI companionship products (Source: techcrunch.com). OpenAI itself acknowledged in the retirement announcement that "losing access to GPT-4o will feel frustrating for some users" and framed its GPT-5.x personality-customization work as a response (Source: openai.com).

Related models

• superseded-by: GPT-5.4 Thinking (by lineage; GPT-4.1 remains in API)
• instance-of: General-Purpose AI (GPAI)
• related: OpenAI, OpenAI o-series (o1 → o4-mini), GPT-5.3 Codex, GPT-5.4 Thinking, GPT-5.5 ('Spud'), gpt-oss (OpenAI open-weight models), (Source: epochai.substack.com), Sparks of Artificial General Intelligence: Early experiments with GPT-4, OpenAI Preparedness Framework V.2, OpenAI Model Spec, Training Language Models to Follow Instructions with Human Feedback (InstructGPT)

The release was accompanied by the GPT-4 System Card, which predates OpenAI's Preparedness Framework. Its best-known element is a preliminary Alignment Research Center evaluation of "GPT-4's ability to carry out actions to autonomously replicate and gather resources — a risk that, while speculative, may become possible with sufficiently advanced AI systems — with the conclusion that the current model is probably not yet capable of autonomously doing so." The card also gives a direct statement of the instrumental-convergence rationale for such testing, and is candid about partial mitigation: "the effectiveness of these mitigations varies," and post-mitigation GPT-4 "is also able to give more detailed guidance on how to conduct harmful or illegal activities" than earlier models.

Relationships

• developer: OpenAI
• superseded-by: GPT-5 Family (OpenAI), GPT-5.4 Thinking, GPT-5.5 ('Spud') (successor OpenAI flagships)
• instance-of: General-Purpose AI (GPAI)
• depends-on: Training Language Models to Follow Instructions with Human Feedback (InstructGPT) (RLHF post-training lineage), Scaling Laws
• supports: Sycophancy and Hallucination (the April 2025 GPT-4o episode), Ask don't tell: Reducing sycophancy in large language models — Dubois, Ududec, Summerfield, Luettgau (UK AISI)
• litigates: NYT v. Microsoft, OpenAI et al., Raine v. OpenAI — Wrongful Death Complaint (2025)
• related: OpenAI o-series (o1 → o4-mini), GPT-5.3 Codex, GPT-5.6 (Sol, Terra, Luna), gpt-oss (OpenAI open-weight models), METR, Anthropic's Responsible Scaling Policy (Version 3.1), OpenAI Preparedness Framework V.2, OpenAI Model Spec, Sparks of Artificial General Intelligence: Early experiments with GPT-4, Microsoft